update and fix conflicts.

.travis.yml

@@ -21,7 +21,6 @@ addons:
       - python
       - python-pip
       - python2.7-dev
-      - python-numpy
       - python-wheel
       - libboost-dev
       - curl
@@ -35,8 +34,8 @@ before_install:
   - if [[ "$JOB" == "check_style" ]]; then sudo ln -s /usr/bin/clang-format-3.8 /usr/bin/clang-format; fi
   # Paddle is using protobuf 3.1 currently. Protobuf 3.2 breaks the compatibility. So we specify the python
   # protobuf version.
-  - pip install -r $TRAVIS_BUILD_DIR/python/requirements.txt
-  - pip install wheel sphinx==1.5.6 recommonmark sphinx-rtd-theme==0.1.9 virtualenv pre-commit LinkChecker
+  - sudo pip install -r $TRAVIS_BUILD_DIR/python/requirements.txt
+  - sudo pip install wheel sphinx==1.5.6 recommonmark sphinx-rtd-theme==0.1.9 virtualenv pre-commit LinkChecker
   - curl https://glide.sh/get | bash
   - eval "$(GIMME_GO_VERSION=1.8.3 gimme)"
   - go get -u github.com/alecthomas/gometalinter

CMakeLists.txt

@@ -65,8 +65,8 @@ if(NOT CMAKE_BUILD_TYPE)
 endif()
 
 if(ANDROID)
-    if(${CMAKE_SYSTEM_VERSION} VERSION_LESS "21")
-        message(FATAL_ERROR "Unsupport standalone toolchains with Android API level lower than 21")
+    if(${CMAKE_SYSTEM_VERSION} VERSION_LESS "16")
+        message(FATAL_ERROR "Unsupport standalone toolchains with Android API level lower than 16")
     endif()
 
     set(WITH_GPU OFF CACHE STRING

doc/design/functions_operators_layers.md

@@ -86,12 +86,13 @@ def layer.fc(X):
 
 We'd like to have Python bindings to operators in package `paddle.operator`, and Python compositions of operators in package `paddle.layer`.  So we have the following concepts in above illustrative example:
 
-```
+
 | C++ functions/functors | mul          | add          |             |          |
+|------------------------|--------------|--------------|-------------|----------|
 | C++ operator class     | mulOp        | addOp        | FCOp        |          |
 | Python binding         | operator.mul | operator.add | operator.fc |          |
 | Python function        |              |              |             | layer.fc |
-```
+
 
 This is how we differentiate layer and operators in PaddlePaddle:
 

doc/design/ops/dist_train.md

@@ -0,0 +1,106 @@
+# Design Doc: Operation Graph Based Parameter Server
+
+## Abstract
+
+We propose an approach to implement the parameter server. In this
+approach, there is no fundamental difference between the trainer and
+the parameter server: they both run subgraphs, but subgraphs of
+different purposes.
+
+## Background
+
+The previous implementations of the parameter server does not run a
+subgraph. parameter initialization, optimizer computation, network
+communication and checkpointing are implemented twice on both the
+trainer and the parameter server.
+
+It would be great if we can write code once and use them on both the
+trainer and the parameter server: reduces code duplication and
+improves extensibility. Given that after the current refactor, we are
+representing everything as a computing graph on the
+trainer. Representing everything as a computing graph on the parameter
+server becomes a natural extension.
+
+## Design
+
+### Graph Converter
+
+The *graph converter* converts the user-defined operation (OP) graph
+into subgraphs to be scheduled on different nodes with the following
+steps:
+
+1. OP placement: the OPs will be placed on different nodes according
+   to heuristic that minimizes estimated total computation
+   time. Currently we will use a simple heuristic that puts parameter
+   varable on parameter server workers and everything else on trainer
+   workers.
+
+1. Add communication OPs to enable the communication between nodes.
+
+We will need these OPs: *Send*, *Recv*, *Enqueue*, *Dequeue*.
+
+Below is an example of converting the user defined graph to the
+subgraphs for the trainer and the parameter server:
+
+<img src="src/local-graph.png" width="300"/>
+
+After converting:
+
+<img src="src/dist-graph.png" width="700"/>
+
+1. The parameter variable W and it's optimizer subgraph are placed on the parameter server.
+1. Operators are added to the subgraphs.
+   - *Send* sends data to the connected *Recv* operator.  The
+	 scheduler on the receive node will only schedule *Recv* operator
+	 to run when the *Send* operator has ran (the *Send* OP will mark
+	 the *Recv* OP runnable automatically).
+   - *Enueue* enqueues the input variable, it can block until space
+     become available in the queue.
+   - *Dequeue* outputs configurable numbers of tensors from the
+     queue. It will block until the queue have the required number of
+     tensors.
+
+
+### Benefits
+
+- Model parallelism become easier to implement: it's an extension to
+  the trainer - parameter server approach. we already have the
+  communication OPs, but need to extend the graph converter's
+  placement functionality.
+
+- User-defined optimizer is easier to add - user can now express it as
+  a subgraph.
+
+- No more duplication logic inside the trainer and the parameter
+  server mentioned in the background section.
+
+### Challenges
+
+- It might be hard for the graph converter to cut a general graph
+  (without any hint for which subgraph is the optimizer). We may need
+  to label which subgraph inside the OP graph is the optimizer.
+
+- It's important to balance the parameter shards of on multiple
+  parameter server. If a single parameter is very big (some
+  word-embedding, fully connected, softmax layer), we need to
+  automatically partition the single parameter onto different
+  parameter servers when possible (only element-wise optimizer depends
+  on the parameter variable).
+
+### Discussion
+
+- In the "Aync SGD" figure, the "W" variable on the parameter server
+  could be read and wrote concurrently, what is our locking strategy?
+  E.g., each variable have a lock cpp method to be invoked by every
+  OP, or, have a lock OP.
+
+- Can the Enqueue OP be implemented under our current tensor design
+  (puts the input tensor into the queue tensor)?
+
+- *Dequeue* OP will have variable numbers of output (depends on the
+  `min_count` attribute), does our current design support it? (similar
+  question for the *Add* OP)
+
+
+### References:
+[1] [TensorFlow: Large-Scale Machine Learning on Heterogeneous Distributed Systems](https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/45166.pdf)

paddle/framework/attribute.h

@@ -45,7 +45,19 @@ class GreaterThanChecker {
  public:
   explicit GreaterThanChecker(T lower_bound) : lower_bound_(lower_bound) {}
   void operator()(T& value) const {
-    PADDLE_ENFORCE(value > lower_bound_, "larger_than check fail");
+    PADDLE_ENFORCE(value > lower_bound_, "larger_than check fails.");
+  }
+
+ private:
+  T lower_bound_;
+};
+
+template <typename T>
+class EqualGreaterThanChecker {
+ public:
+  explicit EqualGreaterThanChecker(T lower_bound) : lower_bound_(lower_bound) {}
+  void operator()(T& value) const {
+    PADDLE_ENFORCE_GE(value, lower_bound_, "equal_larger_than check fails.");
   }
 
  private:
@@ -115,6 +127,11 @@ class TypedAttrChecker {
     return *this;
   }
 
+  TypedAttrChecker& EqualGreaterThan(const T& lower_bound) {
+    value_checkers_.push_back(EqualGreaterThanChecker<T>(lower_bound));
+    return *this;
+  }
+
   // we can add more common limits, like LessThan(), Between()...
 
   TypedAttrChecker& SetDefault(const T& default_value) {

paddle/framework/backward.md

@@ -2,20 +2,20 @@
 
 ## Motivation
 
-In Neural Network, the backpropagation algorithm follows the chain rule, so we need to compound the fundmental gradient operators/expressions together with chain rule . Every forward network need a backward network to construct the full computation graph, the operator/expression's backward pass will be generated respect to forward pass.
+In Neural Network, the backpropagation algorithm follows the chain rule, so we need to compound the gradient operators/expressions together with the chain rule. Every forward network needs a backward network to construct the full computation graph, the operator/expression's backward pass will be generated respect to forward pass.
 
 ## Backward Operator Registry
 
-A backward network is built up with several backward operators. Backward operators take forward operators' inputs, outputs and output gradients and then calculate its input gradients.
+A backward network is built up with several backward operators. Backward operators take forward operators' inputs outputs, and output gradients and then calculate its input gradients.
 
 |                        | forward operator | backward operator 
 | ---------------------- | ---------------- |------------------------- |		
 | **Operator::inputs_**  | Inputs       | Inputs, Outputs, OutputGradients |	
 | **Operator::outputs_** | Outputs          | InputGradients            |
 
- In most cases, there is a one-to-one correspondence between forward and backward operators. These correspondences are recorded by a global hash map(`OpInfoMap`). To follow the philosophy of minimum core and make operators pluggable, the registry mechanism is introduced.
+ In most cases, there is a one-to-one correspondence between the forward and backward operators. These correspondences are recorded by a global hash map(`OpInfoMap`). To follow the philosophy of minimum core and make operators pluggable, the registry mechanism is introduced.
 
-For example, we have got a `mul_op`, and we can register it's information and corresponding backward operator by the following macro:
+For example, we have got a `mul_op`, and we can register its information and corresponding backward operator by the following macro:
 
 ```cpp
 REGISTER_OP(mul, MulOp, MulOpMaker, mul_grad, MulOpGrad);
@@ -27,7 +27,7 @@ REGISTER_OP(mul, MulOp, MulOpMaker, mul_grad, MulOpGrad);
 
 ## Backward Opeartor Creating
 
-Given a certain forward operator, we can get its corresponding backward opeartor by calling:
+Given a certain forward operator, we can get its corresponding backward operator by calling:
 
 ```cpp
 OperatorBase* bwd_op = BuildGradOp(const OperatorBase* fwd_op);
@@ -37,7 +37,7 @@ The function `BuildGradOp` will sequentially execute following processes:
 
 1. Get the `type_` of given forward operator, and then get the corresponding backward operator's type by looking up the `OpInfoMap`.
 
-2. Build two maps named `inputs` and `outputs` to temporary storage backward operator's inputs and outputs. Copy forward operator's `inputs_` and `outputs_` to map `inputs`, except these are not necessary for gradient computing.
+2. Build two maps named `inputs` and `outputs` to temporary storage backward operator's inputs and outputs. Copy forward operator's `inputs_` and `outputs_` to map `inputs`, except these, are not necessary for gradient computing.
 
 3. Add forward inputs' gradient variables into map `output`, adding forward outputs' gradient variables into map `input`.
 
@@ -49,31 +49,31 @@ A backward network is a series of backward operators. The main idea of building
 
 In our design, the network itself is also a kind of operator. So the operators contained by a big network may be some small network. 
 
-given a forward network, it generates the backward network. We only care about the Gradients—`OutputGradients`,`InputGradients`.
+given a forward network, it generates the backward network. We only care about the Gradients—`OutputGradients`, `InputGradients`.
 
 1. Op 
 
-   when the input forward network is a Op, return its gradient Operator Immediately.
+   when the input forward network is an Op, return its gradient Operator Immediately.
 
 2. NetOp 
 
-   when the input forward network is a NetOp, it need to call the sub NetOp/Operators backward function recursively. During the process, we need to collect the `OutputGradients` name according to forward NetOp.
+   when the input forward network is a NetOp, it needs to call the sub NetOp/Operators backward function recursively. During the process, we need to collect the `OutputGradients` name according to the forward NetOp.
 
-   **shared variable**. As illustrated in the pictures, two operator's `Output` `Gradient` will overwirte their shared input variable.  
+   **shared variable**. As illustrated in the pictures, two operator's `Output` `Gradient` will overwrite their shared input variable.  
 
    <p align="center">
-   <img src="./images/duplicate_op.png" width="70%" ><br/>
+   <img src="./images/duplicate_op.png" width="50%" ><br/>
 
-   1. shared variable in two operators. 
+   1. Shared variable in operators. 
 
    </p>
 
-   Share variable between operators or same input variable used in multiple operators lead to a duplicate gradient variable. As demo show above, we need to rename gradient name recursively, and add a generic add operator replace the overwirte links. 
+   Share variable between operators or same input variable used in multiple operators leads to a duplicate gradient variable. As demo show above, we need to rename gradient name recursively and add a generic add operator replace the overwrite links. 
 
    <p align="center">
-   <img src="images/duplicate_op2.png" width="90%" ><br/>
+   <img src="images/duplicate_op2.png" width="50%" ><br/>
 
-   2. replace shared variable gradient with `Add` Operator
+   2. Replace shared variable's gradient with `Add` operator.
 
    </p>
 

paddle/framework/ddim.cc

@@ -283,5 +283,14 @@ std::ostream& operator<<(std::ostream& os, const DDim& ddim) {
 DDim::DDim(std::initializer_list<int64_t> init_list) {
   *this = make_ddim(init_list);
 }
+
+DDim flatten_to_2d(const DDim& src, int num_col_dims) {
+  int rank = src.size();
+  return make_ddim({product(slice_ddim(src, 0, num_col_dims)),
+                    product(slice_ddim(src, num_col_dims, rank))});
+}
+
+DDim flatten_to_1d(const DDim& src) { return make_ddim({product(src)}); }
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/ddim.h

@@ -115,6 +115,12 @@ int arity(const DDim& ddim);
 
 std::ostream& operator<<(std::ostream&, const DDim&);
 
+// Reshape a tensor to a matrix. The matrix's first dimension(column length)
+// will be the product of tensor's first `num_col_dims` dimensions.
+DDim flatten_to_2d(const DDim& src, int num_col_dims);
+
+DDim flatten_to_1d(const DDim& src);
+
 }  // namespace framework
 }  // namespace paddle
 

paddle/framework/eigen.h

@@ -63,20 +63,35 @@ struct EigenTensor {
 
 template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
-struct EigenMatrix : public EigenTensor<T, 2, MajorType, IndexType> {};
+struct EigenMatrix : public EigenTensor<T, 2, MajorType, IndexType> {
+  static typename EigenMatrix::Type Reshape(Tensor& tensor, int num_col_dims) {
+    int rank = tensor.dims_.size();
+    PADDLE_ENFORCE(num_col_dims > 0 && num_col_dims < rank,
+                   "`num_col_dims` must be between (0, rank_of_tensor).");
+    return EigenMatrix::From(tensor,
+                             flatten_to_2d(tensor.dims(), num_col_dims));
+  }
+
+  static typename EigenMatrix::ConstType Reshape(const Tensor& tensor,
+                                                 int num_col_dims) {
+    int rank = tensor.dims_.size();
+    PADDLE_ENFORCE(num_col_dims > 0 && num_col_dims < rank,
+                   "`num_col_dims` must be between (0, rank_of_tensor).");
+    return EigenMatrix::From(tensor,
+                             flatten_to_2d(tensor.dims(), num_col_dims));
+  }
+};
 
 template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 struct EigenVector : public EigenTensor<T, 1, MajorType, IndexType> {
   // Flatten reshapes a Tensor into an EigenVector.
   static typename EigenVector::Type Flatten(Tensor& tensor) {
-    return EigenVector::From(
-        tensor, make_ddim({static_cast<int>(product(tensor.dims_))}));
+    return EigenVector::From(tensor, {product(tensor.dims_)});
   }
 
   static typename EigenVector::ConstType Flatten(const Tensor& tensor) {
-    return EigenVector::From(
-        tensor, make_ddim({static_cast<int>(product(tensor.dims_))}));
+    return EigenVector::From(tensor, {product(tensor.dims_)});
   }
 };
 

paddle/framework/eigen_test.cc

@@ -108,5 +108,24 @@ TEST(Eigen, Matrix) {
   }
 }
 
+TEST(Eigen, MatrixReshape) {
+  Tensor t;
+  float* p = t.mutable_data<float>({2, 3, 6, 4}, platform::CPUPlace());
+  for (int i = 0; i < 2 * 3 * 6 * 4; ++i) {
+    p[i] = static_cast<float>(i);
+  }
+
+  EigenMatrix<float>::Type em = EigenMatrix<float>::Reshape(t, 2);
+
+  ASSERT_EQ(2 * 3, em.dimension(0));
+  ASSERT_EQ(6 * 4, em.dimension(1));
+
+  for (int i = 0; i < 2 * 3; i++) {
+    for (int j = 0; j < 6 * 4; j++) {
+      ASSERT_NEAR(i * 6 * 4 + j, em(i, j), 1e-6f);
+    }
+  }
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/operator.cc

@@ -123,6 +123,15 @@ OperatorBase::OperatorBase(const std::string& type,
   CheckAllInputOutputSet();
 }
 
+std::vector<std::string> OperatorBase::InputVars() const {
+  std::vector<std::string> ret_val;
+  for (auto& o : outputs_) {
+    ret_val.reserve(ret_val.size() + o.second.size());
+    ret_val.insert(ret_val.end(), o.second.begin(), o.second.end());
+  }
+  return ret_val;
+}
+
 std::vector<std::string> OperatorBase::OutputVars(bool has_intermediate) const {
   std::vector<std::string> ret_val;
   if (has_intermediate) {

paddle/framework/operator.h

@@ -94,11 +94,14 @@ class OperatorBase {
 
   const VariableNameMap& Inputs() const { return inputs_; }
   const VariableNameMap& Outputs() const { return outputs_; }
+
   //! Get a input with argument's name described in `op_proto`
   std::string Input(const std::string& name) const;
   //! Get a input which has multiple variables.
   const std::vector<std::string>& Inputs(const std::string& name) const;
 
+  std::vector<std::string> InputVars() const;
+
   //! Get a output with argument's name described in `op_proto`
   std::string Output(const std::string& name) const;
   //! Get an output which has multiple variables.
@@ -311,9 +314,9 @@ class InferShapeContext {
   }
 
   template <typename T>
-  std::vector<const T*> MultiOutput(const std::string& name) const {
+  std::vector<T*> MultiOutput(const std::string& name) const {
     auto names = op_.Outputs(name);
-    std::vector<const T*> res;
+    std::vector<T*> res;
     res.reserve(names.size());
     std::transform(names.begin(), names.end(), std::back_inserter(res),
                    [&](const std::string& sub_name) {

paddle/framework/tensor.h

@@ -43,6 +43,9 @@ class Tensor {
   template <typename T, size_t D, int MajorType, typename IndexType>
   friend struct EigenTensor;
 
+  template <typename T, int MajorType, typename IndexType>
+  friend struct EigenMatrix;
+
   template <typename T, int MajorType, typename IndexType>
   friend struct EigenVector;
 

paddle/framework/tensor_impl.h

@@ -148,5 +148,13 @@ inline Tensor& Tensor::Resize(const DDim& dims) {
 
 inline const DDim& Tensor::dims() const { return dims_; }
 
+template <typename T>
+inline Tensor ReshapeToMatrix(const Tensor& src, int num_col_dims) {
+  Tensor res;
+  res.ShareDataWith<T>(src);
+  res.Resize(flatten_to_2d(src.dims(), num_col_dims));
+  return res;
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/tensor_test.cc

@@ -262,3 +262,16 @@ TEST(Tensor, CopyFrom) {
   }
 #endif
 }
+
+TEST(Tensor, ReshapeToMatrix) {
+  using namespace paddle::framework;
+  using namespace paddle::platform;
+  Tensor src;
+  int* src_ptr = src.mutable_data<int>({2, 3, 4, 9}, CPUPlace());
+  for (int i = 0; i < 2 * 3 * 4 * 9; ++i) {
+    src_ptr[i] = i;
+  }
+  Tensor res = ReshapeToMatrix<int>(src, 2);
+  ASSERT_EQ(res.dims()[0], 2 * 3);
+  ASSERT_EQ(res.dims()[1], 4 * 9);
+}

paddle/gserver/layers/BatchNormBaseLayer.cpp

@@ -62,14 +62,18 @@ void BatchNormBaseLayer::calFeatureMapSize() {
   const ImageConfig& conf = config_.inputs(0).image_conf();
   imageH_ = inputLayers_[0]->getOutput().getFrameHeight();
   imageW_ = inputLayers_[0]->getOutput().getFrameWidth();
+  imageD_ = inputLayers_[0]->getOutput().getFrameDepth();
+
+  if (0 == imageD_) imageD_ = conf.img_size_z();
   if (imageH_ == 0 && imageW_ == 0) {
     imageH_ = conf.has_img_size_y() ? conf.img_size_y() : conf.img_size();
     imageW_ = conf.img_size();
   } else {
     getOutput().setFrameHeight(imageH_);
     getOutput().setFrameWidth(imageW_);
+    getOutput().setFrameDepth(imageD_);
   }
-  imgPixels_ = imageH_ * imageW_;
+  imgPixels_ = imageH_ * imageW_ * imageD_;
 }
 
 }  // namespace paddle

paddle/gserver/layers/BatchNormBaseLayer.h

@@ -80,6 +80,7 @@ protected:
 
   /// Height or width of input image feature.
   /// Both of them are 1 if the input is fully-connected layer.
+  int imageD_;
   int imageH_;
   int imageW_;
   /// Height * Width.

paddle/gserver/layers/CudnnBatchNormLayer.cpp

@@ -37,7 +37,7 @@ bool CudnnBatchNormLayer::init(const LayerMap& layerMap,
 }
 
 void CudnnBatchNormLayer::reshape(int batchSize) {
-  hl_tensor_reshape(ioDesc_, batchSize, channels_, imageH_, imageW_);
+  hl_tensor_reshape(ioDesc_, batchSize, channels_, imageH_ * imageD_, imageW_);
 }
 
 void CudnnBatchNormLayer::forward(PassType passType) {
@@ -104,7 +104,7 @@ void CudnnBatchNormLayer::forward(PassType passType) {
                                    EPS,
                                    batchSize,
                                    channels_,
-                                   imageH_,
+                                   imageH_ * imageD_,
                                    imageW_);
     }
   }

paddle/gserver/layers/SwitchOrderLayer.cpp

@@ -24,19 +24,21 @@ bool SwitchOrderLayer::init(const LayerMap& layerMap,
   /* Initialize the basic parent class */
   Layer::init(layerMap, parameterMap);
   auto& img_conf = config_.inputs(0).image_conf();
+  size_t inD = img_conf.img_size_z();
   size_t inH =
       img_conf.has_img_size_y() ? img_conf.img_size_y() : img_conf.img_size();
   size_t inW = img_conf.img_size();
   size_t inC = img_conf.channels();
+  inH = inH * inD;
   inDims_ = TensorShape({0, inC, inH, inW});
   outDims_ = TensorShape(4);
 
   auto& reshape_conf = config_.reshape_conf();
-  for (size_t i = 0; i < reshape_conf.heightaxis_size(); i++) {
-    heightAxis_.push_back(reshape_conf.heightaxis(i));
+  for (int i = 0; i < reshape_conf.height_axis_size(); i++) {
+    heightAxis_.push_back(reshape_conf.height_axis(i));
   }
-  for (size_t i = 0; i < reshape_conf.widthaxis_size(); i++) {
-    widthAxis_.push_back(reshape_conf.widthaxis(i));
+  for (int i = 0; i < reshape_conf.width_axis_size(); i++) {
+    widthAxis_.push_back(reshape_conf.width_axis(i));
   }
   createFunction(nchw2nhwc_, "NCHW2NHWC", FuncConfig());
   createFunction(nhwc2nchw_, "NHWC2NCHW", FuncConfig());
@@ -64,9 +66,10 @@ void SwitchOrderLayer::setInDims() {
   MatrixPtr input = inputLayers_[0]->getOutputValue();
   size_t batchSize = input->getHeight();
   inDims_.setDim(0, batchSize);
-
+  int d = inputLayers_[0]->getOutput().getFrameDepth();
+  d = (d == 0 ? 1 : d);
   int h = inputLayers_[0]->getOutput().getFrameHeight();
-  if (h != 0) inDims_.setDim(2, h);
+  if (h != 0) inDims_.setDim(2, h * d);
   int w = inputLayers_[0]->getOutput().getFrameWidth();
   if (w != 0) inDims_.setDim(3, w);
   int totalCount = input->getElementCnt();